Known and Unknown Facts of LoRa: Experiences from a …Known and Unknown Facts of LoRa: Experiences from a Large Scale Measurement Study JANSEN C. LIANDO, AMALINDA GAMAGE, AGUSTINUS

Known and Unknown Facts of LoRa: Experiences from aLarge Scale Measurement Study

JANSEN C. LIANDO, AMALINDA GAMAGE, AGUSTINUS W. TENGOURTIUS, and MOLI, School of Computer Science and Engineering, Nanyang Technological University, Singapore

LoRa is an LPWAN technology designed for the IoT. In recent years, it has gained significant momentumamongst both industrial and research communities. Patented by Semtech, LoRa makes use of chirp spreadspectrum modulation to deliver data with promises of long battery life, far-reaching communication distances,and a high node density at the cost of data rate. In this paper we conduct a series of experiments to verify theclaims made by Semtech on LoRa technology. Our results show that LoRa is capable of communicating over10km under line-of-sight environments. However, under non-line-of-sight environments, LoRa’s performanceis severely affected by obstructions such as buildings and vegetations. Moreover, the promise of prolongedbattery life requires extreme tuning of parameters. Lastly, a LoRa gateway supports up to 6,000 nodes withPRR requirement of >70%. This study also explores the relationship between LoRa transmission parametersand proposes an algorithm to determine optimal settings in terms of coverage and power consumption undernon-line-of-sight environments. It further investigates the impact of LoRaWAN on energy consumption andnetwork capacity along with implementation of a LoRa medium access mechanism and possible gains broughtforth by implementing such a mechanism.

Additional Key Words and Phrases: Network measurement, Network performance analysis, Sensor networks,Wide area networks, Low power wide area network, Internet of things, LoRa

ACM Reference Format:Jansen C. Liando, Amalinda Gamage, Agustinus W. Tengourtius, and Mo Li. 2018. Known and UnknownFacts of LoRa: Experiences from a Large Scale Measurement Study. ACM Trans. Sensor Netw. 0, 0, Article 19(November 2018), 35 pages. https://doi.org/0000001.0000001

1 INTRODUCTIONLow Power Wide Area Networks (LPWANs) are garnering increased attention from the researchcommunity and is expected to be the missing piece for the future scalable Internet of Things (IoT).Recent new LPWAN developments are aimed at fulfilling the gaps composed of limitations thatmainly constitute to battery life, coexistence, and communication range of IoT devices.Long Range (LoRa) is a recent industry initiated LPWAN technology by Semtech among these

efforts in building a scalable IoT. LoRa, shortly after its release, captured considerable attentionfrom scientific and industrial communities and brought together multilateral efforts with strongpotential to collaborate on the development of worldwide coverage for IoT devices.

This work is supported by Singapore MOE Tier 2 grant MOE2016-T2-2-023, Tier 1 grant 2017-T1-002-047, and NTU CoEgrant M4081879.Authors’ address: Jansen C. Liando; Amalinda Gamage; Agustinus W. Tengourtius; Mo Li, School of Computer Science andEngineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore, 639798, Singapore, {cjansen,amalinda,atengourtius,limo}@ntu.edu.sg.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without feeprovided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice andthe full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requiresprior specific permission and/or a fee. Request permissions from [email protected].© 2018 Association for Computing Machinery.1550-4859/2018/11-ART19 $15.00https://doi.org/0000001.0000001

ACM Trans. Sensor Netw., Vol. 0, No. 0, Article 19. Publication date: November 2018.

https://doi.org/0000001.0000001

https://doi.org/0000001.0000001

19:2 Jansen C. L. et al.

LoRa’s popularity stems from a range of features and performance promises. Firstly, LoRa utilizesthe unlicensed Industrial, Scientific, and Medical radio (ISM) bands and promises kilometers ofcommunication distance and several years of battery life. Secondly, LoRa incorporates a variationof Chirp Spread Spectrum (CSS) [17] technology that Semtech claims to be robust against a highdegree of interference in addition to multi-path and Doppler effects [2, 34, 41]. Thirdly, LoRa alsoclaims to support high channel capacity, meaning that a single channel supports simultaneoustransmissions from multiple devices which is in high contrast to conventional LPWAN technologies.Although Lora’s physical layer is a proprietary technology acquired by Semtech [51, 52], partsfrom the same is open for public [11, 34]. In this paper, we exploit those public information towiden the existing knowledge on LoRa. Despite clinging to claims printed on data sheets, ourmeasurement study tests such claims with practical conditions and real-life scenarios. We performexperimental investigations to further understand the room of optimizing and improving theLoRa performance in various ways. Our study aims at answering researchers or system adopters’questions like "What are the most energy efficient LoRa configurations under different distanceand environment conditions?", "How the upper LoRaWAN MAC behaviors have an impact on theLoRa system performance in terms of the energy and channel efficiency?", "What is the LoRa PHYefficiency? How can we improve in terms of medium access efficiency, reliability and robustness,etc.?".

We primarily experiment with our campus deployment comprised of handcrafted LoRa gatewaysdeployed on the rooftops of three buildings, a LoRa network server at the cloud side, 50+ endnodes tested with various conditions, a mobile LoRa gateway, as well as various wireless spectrumsensing and data logging devices. The measurement study is performed within an area of 3km×3kmcomprising the university campus and nearby region of residence. This study collected and analyzed>100 sensor traces with >2,000,000 lines of data. In order to reveal the PHY details of LoRa, wefurther build a software defined radio (SDR) enabled LoRa gateway based on USRP N210 and GNURadio, with which we can intercept on-the-air LoRa transmissions and perform PHY level signalanalysis.

This paper is organized as follows. Section 2 briefly introduces fundamentals of LoRa modulationand our experiment setup. Section 3 reports our measurement findings and derived models inaccordance with the performance promises by LoRa. Section 4 exploits the optimization andimprovement opportunities and explores beyond existing scope of LoRa practices by gaininginsights through SDR based LoRa PHY analysis. Section 5 discusses some other LoRa measurementstudies and concludes this paper.

2 OVERVIEW OF LORAWe briefly introduce LoRa modulation and the key technical parameters in order to establish acommon understanding to better comprehend Sections 3 and 4. In LoRa, chirp and symbol bothrefers to the modulated signal which carries data and in this paper, we use them interchangeably.

2.1 LoRa ModulationLoRa employs Chirp Spread Spectrum (CSS) modulation to modulate signals. A chirp in CSS refersto a signal with constantly increasing or decreasing frequency that sweeps through and wrapsaround a predefined bandwidth, referred as upchirps and downchirps. Fig. 1 visualizes a LoRa packetcaptured using a software defined radio setup from our experiment. The figure illustrates differenttypes of chirps - the first half being standard upchirps where frequency increases over time andrestarts from the min frequency (fmin ) towards max frequency (fmax ) followed by short downchirpsannotated as Start Frame Delimiter (SFD) that goes from fmax to fmin , and then modulated chirpsthat contain data bits.


Known and Unknown Facts of LoRa 19:3Fr

eque

ncy

(KH

z)

Time (s)

Preamble SFD Data

Fig. 1. A snapshot of LoRa transmission that shows up, down and data chirps as seen on spectrogram.

The simplest implementation of CSS is on-off keying which modulates data into upchirps anddownchirps. The LoRa implementation of CSS however, differs from that and modulates data throughthe changes in the chirp’s starting frequency position. The instantaneous change in the startingfrequency indicates the symbol border as illustrated in the modulated chirps in Fig. 1.Theoretically, LoRa is able to achieve a data rate up to 27kbit/s. The data rate while limited, is

more than sufficient for LPWAN applications where communication coverage is prioritized overdata rate. LoRa configuration can be modified by manipulating some key parameters to achievetrade-offs among communication distances, data rate, and power consumptions.

2.2 Key ParametersLoRa employs a unique variation of Chirp Spread Spectrum modulation which incorporates fourkey parameters- channel, bandwidth, spreading factor and transmission power. As depicted in Fig.1, LoRa utilizes a combination of upchirps and downchirps to formulate a full packet. The angle ofchirps, known as the spreading factor, remains consistent throughout the packet. For a system withconstant frequency bandwidth, the spreading factor determines the final data rate. LoRa modulation,at the time, supports seven different spreading factors ranging from SF6 to SF12. SF6 having thehighest data rate requires the highest signal-to-noise ratio (SNR) for successful demodulation whileSF12 supports lowest data rate but requires lowest SNR for the same transmission power.

Table 1. Configuration options provided by Semtech LoRa chipset.

Parameter OptionsSF 6, 7, 8, 9, 10, 11, 12BW (kHz) 7.8, 10.4, 15.6, 20.8, 31.2, 41.7, 62.5, 125, 250, 500CR 4/5, 4/6, 4/7, 4/8IH True or FalseDE True or False (recommended True for Tsym > 16ms)CRC True (uplink) or False (downlink)PL 0 - 255 BytesSpre 6 - 65535 symbols

A chirp’s gradient is determined by predefined parameters offered by Semtech’s LoRa transceiverwhich includes the spreading factor (SF) and frequency bandwidth (BW). The relationship of SF andBW is explained in [34] where each chirp consists of 2SF number of RF chips carrying SF numberof data bits. The number of chips that formulate a complete chirp is directly proportional to BWand is computed as a chip per Hz of BW . Therefore, a standard LoRa chirp that sweeps a BW of125KHz comprises of 125,000 chips/sec. The duration Tsym of such a chirp is given in Eq-1.



Tsym =2SFBW

(1)

Preamble PHY Header PHY Header CRC

6 ‐ 65535 symbols 0 – 3 bytes (encoded in CR4/8)

Mandatory Preamble

4.25 symbols

Payload CRC

if Uplinkif Implicit Header disabled

0 – 255 bytes 0 – 2 bytes

Preamble PHY Header PHY Header CRC

6 ‐ 65535 symbols 0 – 3 bytes (encoded in CR4/8)

Mandatory Preamble

4.25 symbols

Payload CRC

if Uplinkif Implicit Header disabled

0 – 255 bytes 0 – 2 bytes

Fig. 2. LoRa PHY packet structure.

In addition to controlling data rate, the diversity of spreading factor choices also increasesco-existence of LoRa devices, i.e., uniquely different chirp gradients (different SFs) allow thedemodulator to develop a factor of resilience against simultaneous transmissions of different SFs inthe same channel and demodulate all of them. Such a feature significantly enhances the multipleaccess efficiency of LoRa which we comprehensively tested and report in Section 3 and Section 4.Semtech LoRa chipset [58] sets a fixed number of options for each parameter as listed in Table

1 along with the recommended guidelines for using each parameter. The most prominent recom-mendation made by Semtech is to use SF settings of SF7 to SF12 and BW 125, 250, and 500kHz[31].Those recommendations were given to ensure the acceptable transmission distance and data ratetrade-off as both SF and BW would affect transmission duration and data rate.The LoRa packet structure is separated into uplink and downlink packets which is indicated by

the presence and absence of packet cyclic redundancy check (CRC). The number of symbols to betransmitted in a packet is heavily dependent on SF except for the set number of preamble symbols(Spre ) and start frame delimiter (SFD) symbols. A LoRa packet consists of Spre preamble symbols,2 mandatory sync word symbols, 2.25 SFD symbols and based on SF, a variable number of datachirps.

For LoRa physical header and payload, LoRa chirps imposes another set of parameters includingcode rate (CR), implicit header (IH), low data rate optimization enabled (DE), and existence of CRCon top of SF. CR determines error correction capability which is based on Hamming code. The CRsetting is represented as 4/x where it indicates 4 information bits padded with x − 4 number ofparity bits. As illustrated in Fig. 2, The PHY header can be excluded by enabling IH (setting IH to1). On the other hand, the DE parameter adds another layer of redundancy to the data bits for anyLoRa transmission when Tsym exceeds 16ms. The existence of CRC at the end of the packet canbe controlled through the CRC flag and is usually modified to differentiate between uplink anddownlink packets. LoRa packet size is limited to 255 Bytes which can be indicated by the singleByte payload size (bpl ) field in the PHY header.

2.3 Experiment SetupThe system setup incorporates three LoRa gateways located within the area of 1km × 1km on topof three buildings on campus (the deployments indicated in Fig. 4). Each gateway is handcraftedand built with a Raspberry-Pi and an IMST IC880A [23] which works as a radio frontend similarto the design in [7]. The antenna connected to the gateway has 4dBi gain and the Raspberry-Piexecutes a single thread of either the LoRa Packet Forwarder or LoRa Packet Logger as required bydifferent experiment purposes [10, 16]. A picture of the gateway hardware is shown in Fig. 3a.More than 50 LoRa end nodes similar to the design found in [27] are utilized throughout the

experiment and tested over an area of over 3km×3km comprising the university campus and nearbyregion of residences. Each node is composed of an Arduino Pro Mini (ATMEGA328P) working as the


Known and Unknown Facts of LoRa 19:5

Fiber-optic Converter

DC Power

Antenna

OpticCable

SX1301

Raspberry Pi + IC880a

(a) The handcrafted LoRa gatewaydeveloped with IC880A frontendand Raspberry-Pi enabled process-ing unit.

Antenna

Li-PoBattery

SX1276

ATMEGA328P

(b) The LoRa end node developedwith SX1276 frontend (inAir9B[24]) and ATMEGA328P process-ing unit.

(c) The software defined radiobased LoRa gateway developedwith USRP N210 and RFX900daughterboard.

Fig. 3. Hardware utilised for the experimental system.

processing unit interfaced with SX1276 working as the radio frontend. Unnecessary componentssuch as voltage regulators and LEDs were taken off the end nodes to ensure high accuracy of powermeasurements. An image of such an end node is given in Fig. 3b. Power profiling for end nodes isachieved using the Monsoon Power Monitor [46] that measures the power at an accuracy of 106samples/sec. We also implement the LoRa gateway functionality with software defined radio (basedon USRP N210 [61] and GNU Radio [49]), and utilize the SDR gateway to intercept on-the-air LoRatransmissions and perform PHY level signal analysis. Fig. 3c depicts the SDR gateway settings.

500m

Fig. 4. Gateway locations placed within the campus grounds.

3 PROMISES OF LORAThis section reports our experiment results to verify the performances promised by LoRa. Wefurther derived models to describe LoRa’s performance by varying existing parameters providedby Semtech chipsets.



3.1 Communication DistanceOne of the key factors in enabling scalable IoT system is the communication distance of end devicesand in this case, Semtech has promised an ultra-long range connectivity for LoRa devices [58]. Inthis section, we would like to quantify the ultra-long range of LoRa devices in various environments.

Parameter ImpactLoRa transmission distance could be affected by the underlying parameter selection. In order tounderstand the impact of each parameter, we first need to understand the effect of each parameterin shaping the resulting wireless signals. There are four main parameters with pronounced effecton LoRa signals. These parameters are listed below:

3.1.1 Spreading Factor (SF).Spreading Factor in LoRa refers to a value which determines how spread out the chirp wouldbe. This level of spread-ness is mainly due to the number of bits crammed into a single chirp.For example, SF7 means that each chirp represents seven bits. As LoRa chirp is modulated byvarying the starting/ending frequency of each chirp, there has to be 2SF number of starting/endingfrequency position. Recall in Section 2, a chirp has to sweep through a given bandwidth and the timeduration for a chirp can be determined through Eq-1 where the time taken for a chirp frequencyto increase/decrease from f iSF to f i+1SF or f i−1SF is 1

BW secs and this is referred to as a chip in [34]. Achirp is formed by going through all possible fSF which results in 2SF number of chips. In this case,with every increment of SF, the time taken to transmit a chirp is effectively doubled assuming allthe other parameters remains the same. By doubling the chirp duration, receivers would have moreopportunities to sample the signal power which results in higher signal-to-noise ratio (SNR). Sincethe definition of SNR is Psiдnal

Pnoise, the higher the signal power Psiдnal as compared to noise power

Pnoise is, the higher the probability of each chirp to be received correctly.

3.1.2 Bandwidth (BW).Bandwidth on the other hand determines the width of the transmitted signal and according to Eq-1,BW would also determine chirp duration. Remember that a chirp consists of 2SF number of chipsand the duration of each chip is 1

BW . By changing the BW, chip duration would change accordinglywhich in turn affects the chirp duration and finally the SNR or that particular chirp. LoRa gatewaychipset are optimized to receive transmission of 125kHz with added option in receiving transmissionof a fixed SF with varying BW (250kHz and 500kHz) and GFSK (Gaussian Frequency Shift Keying)transmissions.

3.1.3 Transmission Power (TX Pow).Transmission power directly affects the amount of power used to transmit a chirp. By increasingTX Pow, the signal will have higher chances of surviving attenuation caused by the environmentwhich effectively increases the signal power Psiдnal received by receivers. Semtech gateway chipsetSX1301 [59] claims a -142dBm sensitivity and is capable of operating even with negative SNR of upto 9dB which means it is able to operate even if the transmitted signal is below the noise floor.

3.1.4 Code Rate (CR).In LoRa, Code Rate refers to the forward error correction code added to a packet before transmission.LoRa utilizes HammingCode as the forward error correction code used in CR. The settings embeddedinto the chips are shown in Table 1. CR setting of 4/5 indicates that with every four bits of data,one bit of correction code will be added and subsequently CR4/8 refers to four bits of data withfour additional correction bits. CR induces overhead to the transmission by increasing the number



of bits to be transmitted. This overhead allows receiver to check for the correctness of the receivedchirps and provides a possibility to correct some erroneous bits from a chirp.

Experiment SetupAs Semtech gateway chipset allows changes of SF, TX Pow, CR on the fly but requires BW to be setto 125kHz, we conducted an experiment to measure the communication distance of each SF, TXPow, and CR (excluding BW) and presents the highest packet reception rate (PRR) among the TXPow and CR for each SF in Fig. 5. With high number of buildings with heights ≥100m blockingline-of-sight (LoS) across Singapore and tallest buildings being private properties with restrictedaccess to public, we performed the LoS experiment by the beach where LoS could be observed at alltime with limited building in the vicinity. The packets with payload of 10 Bytes were transmitted byseveral end nodes across the 9.08km stretch while maintaining LoS at all times. Each node transmitsequal number of packets for every combination of settings.

0 5 10 15

Distance (KM)

0

0.2

0.4

0.6

0.8

1

Pac

ket R

ecep

tion

Rat

io

SF7SF8SF9SF10SF11SF12

(a) Line-of-Sight.

0 1 2 3 4 5 6 7

Distance (KM)

0

0.2

0.4

0.6

0.8

1P

acke

t Rec

eptio

n R

atio


(b) Non-Line-of-Sight.

Fig. 5. Communication distance of each Spreading Factor on different environment. LoRa devices are stillable to communicate beyond 9km for line-of-sight environment however, LoRa signal propagation is severelyimpacted when obstructed.

For the non-line-of-sight (NLoS) experiment, we performed the experiment in our campus groundwhich represents a metropolitan city scape with high density of high-rise buildings (buildingsheight of ≥100m on average) blocking LoS between transmitter and receiver. The experiments wereconducted on a 3km × 3km area with high density of high-rise structures and natural vegetations.In this experiment, end nodes transmit packets of different settings with payload of 10 Bytescontinuously for ten minutes. Equal number of packets were sent for each setting from a singlenode. The end nodes were placed under open-air with no shelter. Buildings and trees would bepresent on test locations obscuring LoS without blocking the sky view.To further test LoRa coverage, similar experiment was conducted for indoor and semi-indoor

environment. Identical to the experiment done for the outdoor environment, each end nodestransmit for ten minutes period with payload of 10 Bytes and each setting will be transmitted equalnumber of times. However, three gateways were placed surrounding the 1km × 1km area wereused to provide a better coverage within the area due to high attenuation caused by buildings.In this experiment, the indoor environment refers to end nodes being placed inside a buildingand within a room while the semi-indoor environment refers to end nodes placed under a shelterwhich has access to open space such as sheltered walkways and corridors by a building. Indoorand semi-indoor environment were combined to provide a view of LoRa coverage for deploymentwithin and around a complex of buildings.



Line-of-SightThe experiment results are presented in Fig. 5a. LoRa is capable of communicating up to 4kmand 5km for PRR 90% and 70% respectively by using the fastest SF setting of SF7. However, theexperiment was unable to uncover the maximum coverage of LoRa transmission beyond SF7.Regression were performed on available data to extrapolate the PRR for distance beyond 9km. Withthe extrapolated data, LoRa is expected to be able to support up to 10km using SF12 with PRR 70%.

Outdoor Environment

(a) Outdoor coverage with TX Pow20dBm and varying SF.

(b) Outdoor coverage with SF12and varying TX Pow.

(c) Indoor coverage with TXPow 20dBm and varying SF.

Fig. 6. LoRa coverage for different environment with PRR >70% and CR4/8. LoRa coverage are severelyimpacted on the east side of the map due to high density of buildings obstructing LoS.

Data presented in Fig. 5b reveals that LoRa performance declined drastically in NLoS scenariofor outdoor urban environment. In such complex environment, LoRa is only capable of coveringthe distance of 0.1km and 0.3km for PRR 90% and 70% respectively using SF7. Similarly, for SF12,< 2km is the limit for PRR 70%. Moreover, similar to the studies presented in [35], the results in Fig.6a indicates that LoRa transmission is heavily impacted by buildings. The transmission coveragewas reduced significantly on the east part of the map as compared to the north and south parts. Thewest part of the map is inaccessible and thus unable to be explored for experiments. As comparedto the line-of-sight coverage, the impact of the environment could reduce the coverage by 82% onaverage and around 90% for the worst case (east side of the map using SF12). This proves that LoRasignals are severely hampered by obstructions such as high-rise buildings. Furthermore, as seenin Fig. 6b, the benefit of increasing the TX Pow is limited and terrain dependent. The increase oftransmission distance on sparse terrain could go up to 30% (16dBm to 20dBm on southwestern sideof the map) however, the average gain in increasing TX Pow is <7%.

Indoor and Semi-Indoor EnvironmentShown in Fig. 6c, the experiment was conducted only within the 2km × 2km area with the highestand lowest points being 77m and 8m respectively. The results are confined within this area whilecoverage beyond this area has been covered in LoS and NLoS experiments. The coverage of lowerSF (SF8 & SF10) forms islands due to obstructions of terrains and buildings blocking the directpaths resulting in high signal attenuation. However, SF12 is capable of covering the entire test areawith little degradation.



Through a series of experiments, LoRa proves to be capable of communicating beyond what wasclaimed by Semtech in LoS environment. However, LoRa coverage degrades significantly with thepresence of obstructions. For optimal coverage, system adopters would have to carefully choose thelocation for gateways and end nodes to ensure LoS. Where LoS is not attainable, system adopterswould have to carefully choose the settings to ensure coverage while ensuring the lifetime of anend node.

3.2 Lifetime of NodesThis section conducts a group of experiments to capture the energy utilization of both the mi-crocontroller (MCU) and LoRa transceiver of an end node under strategic conditions to exploreuseful information. These experiments are guided by a variance of SF, CR, TX Pow and BW. Usefulcombinations of parameters were selected to conduct experiments for analyzing important aspectsof energy consumption of both end nodes MCU and LoRa transceivers. Firstly, we state the generictransmission energy profile of the end node for a single transmission. This energy profile is obtainedusing Monsoon Power Monitor and readings can be exported for further processing. Fig. 7 illustratesthe data exported from the power monitor. Energy consumption of a single transmission can bebroken down to several components. These components can be grouped into active and sleep states.In active state, there are operating MCU which requires energy to execute tasks and also Radiowhich also require energy to transmit the packet. For sleep state, MCU still requires energy fortime counting. Although small, the energy could take up to a considerable amount across a longperiod of time. Monsoon Power Monitor is used throughout this section to obtain ground truth ofthe hardware’s energy profile.

0.01 0.02 0.03 0.04 0.05

Time (s)

0

50

100

150

Pow

er (

mW

)

Sleep Sleep

TX

MCU Active

Fig. 7. Energy consumption captured by power monitor for a LoRa transmission with parameter SF7, CR4/5,125kHz BW, 2dBm TX Pow, and 9 Bytes payload. Several states could be observed and disected to model theenergy consumption of a LoRa end node.

The first experiment was conducted to compute the energy per transmission for all SFs rangingfrom SF7 to SF12 with a payload of 6 Bytes. Energy profile of the MCU and the transceiver wereseparately recorded for each transmission and is presented in Table 2. This experiment presentsthe two extremes in LoRa transmission configuration. The SF7 2dBm setting representing theleast energy consuming configuration and the SF12 20dBm setting representing the most energyconsuming parameter configuration. With 15 minutes interval between each packet, the experimentshows that LoRa nodes can sustain a lifetime of 4.60 years and 1.37 years respectively. These indicates



that the 10 years end node lifetime promised by LoRa would be possible only through carefulselection of parameter configuration and duty cycling.

Table 2. LoRa packets energy budget breakdown with PL = 6 Bytes, CR = 4/8, BW = 125kHz, 15 minutes perpacket for SF7 & SF12, and battery capacity of 3.7V 2Ah.

States Time (ms) Energy (mJ) Budget (%)

SF72dBm

MCU Active 40.50 0.50 0.30MCU Sleep 899959.50 71.28 43.14Radio TX 38.85 4.36 2.64Radio Sleep 899961.15 89.10 53.92

Total 165.24 4.60 years

SF1220dBm

MCU Active 933.00 12.25 2.22MCU Sleep 899067.00 71.21 12.87Radio TX 926.70 380.73 68.82Radio Sleep 899071.30 89.01 16.09

Total 553.20 1.37 years

The results in Table 2 can be calculated using Eq-2. The parameter Tcycle refers to the timeduration of a single transmission cycle with duty cycle constraint. Energy parameters Ebatt andEcycle each refers to the energy contained in a particular battery and energy expended for eachTcycle respectively. The lifetime calculated in Eq-2 assumes a perfect battery with no degradationdue to time or environmental influences.

Li f etime = Tcycle ×EbattEcycle

(2)

We compute the time duration for a transmission cycle as a function of duty cycle in Eq-3. Asduty cycle is the percentage of time a radio is transmitting, we compute the cycle duration Tcyclewith duty_cycle as the indication of a fraction of time the radio is allowed to transmit where Tpktrepresents the transmission duration of a packet.

Tcycle = 100 ×Tpkt

duty_cycle (3)

The energy contained in a particular battery can be obtained through Eq-4 where the constantvalue of 3600 refers to the number of seconds in an hour, Cbatt refers to the charge in the batterywith the unit of Ampere per hour (Ah), and Vnom as the nominal voltage of the battery.

Ebatt = 3600 ×Cbatt ×Vnom (4)Recall in Fig. 2 where LoRa PHY packet structure contains mandatory preamble of 2 symbols of

sync word and 2.25 symbols of SFD, the transmission duration (Tpkt ) can be calculated using Eq-5with the number of preamble symbols defined by user (Spre ), number of payload symbol (Spl ), andTsym defined in Eq-1.

Tpkt = Tsym(Spre + 2 + 2.25 + Spl ) (5)In order to obtain Ecycle , on top of Tpkt , the energy expended by radio and MCU is required.

These energy levels will have to be separated into two main states, sleep and active/TX.



Table 3. Energy consumption for different type of MCU during a LoRa packet transmission.

MCU EMCU _On EMCU _Of f

Arduino Pro Mini 12.49mW 81.08µWArduino Uno 23.48mW 174.65µWRaspberry Pi 1.41W

Table 3 presents the energy consumption by MCU for both the active (EMCU _On) and sleep(EMCU _Of f ) states. The Ecycle mentioned in Eq-2 can be obtained through Eq-6 where Ecycle isdivided into two parts, the energy expended during sleep state and energy expended during activestate which the duration is based on Tcycle − Tpkt and Tpkt respectively. The energy expendedduring sleep state requires the information of EMCU _Of f and also energy expended by Radio duringsleep ER_Of f while the energy expended during transmission would require the information ofEMCU _On and energy expended by Radio during transmission ER_TX .

Ecycle =[ (Tcycle −Tpkt

) (EMCU _Of f + ER_Of f

) ]+[Tpkt

(ER_TX + EMCU _On

) ](6)

Semtech datasheet provides Eq-7 to calculate Spl . The variables in the equation are Payload inbytes (PL), SF, CR, IH as 0 or 1 to indicate explicit or implicit header respectively, and both CRCand DE as 0 or 1 to indicate absence or presence as required. DE is recommended to be set to 1 ifTsym exceeds 16ms .

Spl = 8 +max

(4CR

⌈8PL − 4SF + 28 + 16CRC − 20IH

4(SF − 2DE)

⌉, 0)

(7)

In order to predict ER_TX we conduct exhaustive tests for all settings that includes SF, BW, CR,TX Power on Arduino Pro Mini with SX1276 (on inAir9B) and RFM96 chipsets. The experimentresulted in >12GByte of measurement data.

2 4 6 8 10 12 14 16 18 20

Transmission Power (dBm)

0

0.1

0.2

0.3

0.4

0.5

Pow

er C

onsu

mpt

ion

(Wat

t)

RFM96SX1276

Fig. 8. ER_TX of different TX Pow across all settings of SF, BW, and CR using Semtech SX1276 and HopeRFRFM96 chipsets. The two chipsets are relatively similar in their power profiles.

We extract the energy consumed by each measurement by converting the energy reading to thescale of Joule per second (Watt). The energy consumption is further grouped by chipsets and TXPow. The values are then averaged and presented in Fig. 8. We compute a cosine similarity for thepower profile of both chipsets and obtained a similarity value of 0.994.



Based on the results in Fig. 8, a polynomial function presented in Eq-8 can be obtained throughregression. We select degree (d) 12 and 10 for SX12761 and RFM962 chipsets respectively for optimalperformance.

ER_TX =d∑n=0

αnxn (8)

2 5 9 13 17 20

Transmission Power (dBm)

100

102

Ene

rgy

per

Pac

ket (

mJ)

SF7 SF9 SF10 SF12

(a) Varying TX Pow.

7 8 9 10 11 12

Spreading Factor (SF)

100

102

Ene

rgy

per

Pac

ket(

mJ)

2dBm 8dBm 14dBm 20dBm

(b) Varying SF.

Fig. 9. Measured and predicted energy consumption with varying settings represented by markers andlines respectively. The model is able to predict the energy consumption of different settings with minimumdeviation.

The results of our prediction model is shown in Fig. 9 where each marker represents actualmeasured data for the particular settings and the line represents the output of prediction model.The energy consumption prediction model achieved a mean error of 0.30% for both chipsets.

Table 4. Predicted nodes lifetime and expected error under specific settings and hardware on a 2Ah battery. Acareful selection of MCU could drastically affect the lifetime of end nodes.

Settings ProMini + RFM96 ProMini + SX1276 Uno + RFM96 Uno + SX1276SF7, CR4/5

125kHz, 10dBm 4.54 years 3.87 years 3.01 years 2.70 yearspayload 10 Bytes ±5 days ±4 days ±3 days ±3 days0.01% duty cycle

SF9 ,CR4/7125kHz, 14dBm 1.71 years 1.65 years 1.41 years 1.37 yearspayload 18 Bytes ±2 days ±2 days ±2 days ±2 days0.1% duty cycleSF12, CR4/8

125kHz, 14dBm 1.42 years 1.39 years 1.21 years 1.19 yearspayload 24 Bytes ±3 days ±2 days ±2 days ±2 days0.1% duty cycle

1RTX coefficients for SX1276 chipset: α0 = -1.860e-01, α1 = 6.258e-01, α2 = -5.478e-01, α3 = 2.724e-01, α4 = -8.534e-02, α5 =1.788e-02, α6 = -2.584e-03, α7 = 2.609e-04, α8 = -1.833e-05, α9 = 8.779e-07, α10 = -2.726e-08, α11 = 4.951e-10, α12 = -3.980e-12.2RTX coefficients for RFM96 chipsets: α0 = -3.523e-01, α1 = 7.024e-01, α2 = -4.630e-01, α3 = 1.680e-01, α4 = -3.724e-02, α5 =5.315e-03, α6 = -4.976e-04, α7 = 3.033e-05, α8 = -1.157e-06, α9 = 2.501e-08, α10 = -2.3351e-10.



We present Table 4 to demonstrate the influence of hardware and setting selections on thelifetime of the end nodes. As expected, duty cycle is the major factor that determines the lifetimeof nodes as the sleep state greatly reduces the energy consumption and allows longer lifetime. SFand TX Pow also plays a large part in the lifetime of end nodes as these settings will determine thetime duration of a packet and the amount of energy to be used for the transmission.

2 5 8 12 15 18 20

TX power

(8,4/5)(8,4/8)(9,4/5)(9,4/8)

(10,4/5)(10,4/8)(11,4/5)(11,4/8)(12,4/5)(12,4/8)

(SF

,CR

)

1.4

1.6

1.8

2

2.2

(a) ER_TX

2 5 8 12 15 18 20

TX power

(8,4/5)(8,4/8)(9,4/5)(9,4/8)

(10,4/5)(10,4/8)(11,4/5)(11,4/8)(12,4/5)(12,4/8)

(SF

,CR

)

0.2

0.3

0.4

(b) Ecycle

Fig. 10. Prediction error for ER_TX and energy consumption of MCU + Radio scenario for 15 minutes inter-packet delay compared to measured readings. By Including MCUs energy consumption, prediction errors ofER_TX become insignificant.

Fig. 10 shows the prediction error of the prediction model for both ER_TX and energy expendedby Ecycle with Tcycle of 15 minutes. Although there is more error for the lower TX Pow settingsin Fig. 10a, the error is not reflected for the same setting in Fig. 10b. This is accredited to thelong duration of the sleep state which contributes to the increase of the proportion for EMCU _Of fand ER_Of f in the Ecycle and dwarfs the prediction error of ER_TX . On the other hand, Fig. 10bshows the prediction error worsen as TX Pow increases. This is due to the hardware variationas the prediction model inputs are the averaged readings of the energy expended across multiplehardware.

0.270.57

1.06

1.66

3.08

5.66

Error Percentage (%)

0

0.5

0.91

Fra

ctio

ns o

f Pac

kets

ER_TX

Ecycle

Fig. 11. CDF of the prediction error per cycle for combinations of SX1276 and RFM96 chipsets with ArduinoUno and Arduino Pro Mini. Regardless of the type of chipsets, prediction errors are still below 6% and withMCUs energy consumption, prediction errors drops to a little over 1%.

Fig. 11 presents the CDF of prediction error as compared to ground truth. The model yields a3.08% error at the 90-th percentile and 1.66% error at the 50-th percentile. As the model only predictsER_TX , researchers and system adopters would be able to easily compute Ecycle by substitutingEMCU _Of f , EMCU _On , and ER_Of f to predict a node energy consumption and lifetime with othertype of MCUs.



3.3 Multiple Access

I/Q

IQ B

ased

RF

Tran

cie

ver

(G)FSK/LoRA Modulators

Packet Handler

Host

Em

bed

ded

Syste

m

SPI

Packet

Han

dle

r

Con

trol8X LoRa

Demodulators

Single CH Single SFLoRa BackhaulDemodulator

(G)FSK Demodulator

SX1301

I/Q

SX

125x

SX

125x

IQ B

ased

RF

Tran

cie

ver

Fig. 12. Block level overview of a LoRa SX1301 concentrator.

This section characterizes LoRa’s Multiple Access performance through a series of experiments.In contrast to common LPWAN protocols such as Sigfox [54], WiFi HaLow [3], LTE-M [1] andWeightless [64], a fundamental difference in LoRa modulation is its capability to perform concurrentcommunications within a single physical channel. This provides LoRa networks a significantadvantage under dense environmental conditions.

This concurrency is due to two factors. Firstly, a standard LoRa gateway receives simultaneoustransmissions from multiple physical channels. Secondly, it also efficiently differentiates betweensimultaneous orthogonal transmissions, even if such transmissions are within the same physicalchannel. In contrast to the capabilities of end nodes based on the SX127X LoRa transceivers, astandard LoRa gateway gains this extra performance at the cost of increased processing power,additional hardware, cost, space and power. Fig. 12 presents an abstract baseband block diagramof a standard LoRa gateway, namely, the IMST IC880A concentrator. The IMST IC880A gatewaydesign incorporates a single SX1301 LoRa baseband signal processor and two full-duplex SX1257 RFfront-end to digital IQ modulator/demodulators. In order to adhere the LoRaWAN standard whichmandates several physical channels based on the ISM band, both front end ICs collaborativelyshare the bandwidth. The SX1301 continuously processes two simultaneous IQ streams from theSX1257 ICs and interprets IQ data in accordance with the LoRa PHY. The decoded information islater made accessible to a host via the SPI protocol.The SX1301 integrates the LoRa IP and supports 10 configurable receiving channels. These 10

falls under three categories based on their use cases and programmability. Eight of them are bydesign fixed in BW to 125KHz while the channel frequencies of those are programmable. They areby design expected to be used to receive packets from end nodes. Out of the remaining two, oneis connected to a GFSK demodulator while the other to a functionally limited LoRa demodulatordesigned to be used as a backhaul channel to other gateways or infrastructure. The limitation on thebackhaul LoRa channel is that it only operates on a single SF of choice as opposed to aforementioned8 LoRa channels which demodulate any given SF. However, the backhaul channel has the advantageof operating on either 125KHz, 250KHz or 500KHz of channel BW.

Semtech asserts that a SX1301 based gateway is able to simultaneously demodulate a maximumof eight concurrent packets [59] as long as non-orthogonal transmissions are separated to differentchannels. Orthogonal transmissions are a key feature in a LoRa modulation that diversify spectrumusage and minimize collisions. They are further explained under Section 4. Given this resiliency toorthogonal transmissions, an estimate of concurrent reception capability at the gateway directly



translates to an indicator that reflects LoRa’s multiple access capability. We leverage this to estimateLoRa’s multiple access capability through a series of experiments.

Capacity of LoRa Spreading FactorsThe experiment setup consisted of 50 randomly placed SX1276 based nodes with payload of 10Bytes and a single IMST IC880A gateway. The 50 nodes were used to emulate a higher number ofnodes based on their duty cycle. The emulation assumed that a typical LoRaWAN node adheres aduty cycle of 0.1% as a viable duty cycle enlisted in the LoRaWAN specification. Therefore, a nodethat transmitted at a duty cycle of 1% was considered to emulate 10 nodes etc.

Fig. 13. Single channel capacity of a SF. Higher SF proves to support more nodes as packet collision reducesas compared to lower SF however, the low throughput of higher SF on contrary to the higher nodes supportedis due to the lower number of packets transmitted.

First, we performed a test to determine the performance of LoRa’s individual SFs. To do so, allnodes were programmed to randomly transmit packets of same SF for a duration 15 minutes. Theduty cycle of real-nodes was controlled to emulate up to 12,000 nodes in steps of 100 under differentexperiment iterations. For each iteration, the PRR and throughput were computed. Fig. 13 depictsthe results of this experiment for selected PRRs of 90%, 80%, 70% and 50%. Based on our results,we also computed the impact on PRR with increasing number of transmitted packets during theperiod of 15 minutes. We present this result in Fig. 14 for all SFs.Although the airtime of SF7 is lowest for a given payload, one may assume that SF7 should

in turn support the largest number of network capacity. However, the result indicated in Fig. 13suggests otherwise. It indicates that SF7 supports the least network capacity and SF12 the highest.Reason for this observation, although counter intuitive, is due to the permitted airtime of the nodesbased on the duty cycle of 0.1%. To elaborate, in comparison to an SF12 end node, an SF7 based endnode is allowed to transmit a higher number of packets as the airtime consumed by same payloadis significantly lower. Effectively, an SF12 based network is permitted to transmit significantly lessnumber of packets than a SF7 based network.The higher number of packets from SF7 end nodes increase the chances of packet loss as a

large number of non-orthogonal transmissions use the same physical channel. This in turn limits



500 1000 2000 4000 8000 16000 32000 64000

#packets

0

0.2

0.4

0.6

0.8

1

PR

R


Fig. 14. PRR of each SF for increasing transmissions. The PRR decreases as the number of packets transmittedincreases which provides a hint in explaining the phenomena observed in Fig. 13.

the concurrent demodulation capability at the gateway. With increased collisions, the number ofcorrectly received packets at the gateway is reduced. Therefore, the number of supported nodesis also reduced. By extension, a SF12 setting supports more nodes achieving a higher networkcapacity under the same duty cycle. In addition, SF7 achieves multiple orders of higher throughputin comparison to a SF12 based network due to the higher numbers of packets transmitted. This isobserved in Fig. 13.

Capacity of a Single LoRa ChannelIn this experiment, we evaluate the capacity of a single LoRa channel as well as the extent ofconcurrent demodulation capability of a LoRa gateway under three different ratios of SFs. Testingthe decoding capability of the concentrator based on the performance of individual SFs is not a trueindicator of the overall performance because a typical LoRa environment contains packets from arange of SFs that arrive at the concentrator. Therefore, a more meaningful method of validatingmultiple access performance of a single LoRa gateway would be to test its performance undercommonly available environmental conditions.

f (β, SF ) = 2 |sf −β | (9)The choice for a SF is mostly motivated by the distance and energy budget in typical LoRa based

applications. For the sake of simplicity, we consider that these applications fall within the threeenvironments namely, indoor, semi-indoor and outdoor. We assume that an indoor environmentconsists more nodes that transmit using lower SFs and vice-versa for an outdoor environment. Wealso assume that a semi-indoor environment consists of nodes with a balanced ratio from all SFs.Having defined the three environments, we now parametrize them in Eq-9. Three values for β inEq-9 are strategically chosen such that the results represent the percentage of nodes for each SF fora fixed value of β , i.e. a set environment. In this case, β = 7 signifies an outdoor network, β = 12 anindoor network while β = SF signifies a semi-indoor network. For further clarity, we present theseratios in Table 5.

Table 5. Experiment settings for single channel performance. An end node in an indoor environment (β = 12)will have a higher chance to transmit in SF7 resulting a higher number of SF7 packets and vice-versa.

Parameter SF7 SF8 SF9 SF10 SF11 SF12β = 12 50.79% 25.40% 12.70% 6.35% 3.17% 1.59%β = SF 16.67% 16.67% 16.67% 16.67% 16.67% 16.67%β = 7 1.59% 3.17% 6.35% 12.07% 25.40% 50.79%



The setup for this experiment is as follows.Firstly, all randomly placed nodes with a pay-load of 10 Bytes were assigned a spreading fac-tor based on the SF ratio for a single β valuerepresenting a particular environment throughEq-9. In contrast to the previous experiment, thisexperiment achieved the emulation of nodes dif-ferently. Here, we assumed a network of LoRanodes that transmitted every 15 minutes. There-fore, a node that transmitted 100 times within15 minutes emulated 100 LoRa nodes etc. Lastly,the aforementioned process was repeated for theremaining two β values.

The results were then processed to determinethe number of nodes that can be supported underPRRs ranging from 50%, 60%, 70%, 80%, and 90%along with respective throughput and goodputrates as presented in Fig. 15. The results of theexperiment indicate that an indoor environment(β = 12) supports the highest throughput andPRR settings.

(a) number of nodes and throughput supported by asingle LoRa channel for varying PRR.

(b) number of nodes and goodput supported by asingle LoRa channel for varying PRR.

Fig. 15. Capacity of a single LoRa channel under mul-tiple β settings. Number of nodes supported, through-put, and goodput increases along with the increaseof SF7 packets.

Fig. 16. Emulated traffic from 2000 nodes received bygateway with β = SF .



This result is intuitive as node density is solely determined by the number of transmissionsunder 15 minutes. In other words, as transmissions that consume less airtime are present in theenvironment, more throughput and a higher number of nodes can be supported for the same PRR.Finally, in order to better present the gateway’s decoding capability, we also present Fig. 16 whichdepicts a snapshot of the decoding progress of the gateway while 10 nodes randomly transmittedpackets using different SF.On the contrary, results in Fig. 15 seemingly disagrees with those in Fig. 13. The results from

our previous experiment that concluded the capacity of SFs favored SF12 nodes. However, inthis experiment, we observed that a higher number of SF7 nodes correspond to an increase inthe total supported nodes. While the results seem contradicting, the latter experiment utilized aratio of different SFs whereas the former utilized a single SF. Recall the condition for concurrentdemodulation of a gateway. The multiple SFs utilized in the latter experiment enhanced theconcurrent decoding capability of the gateway while the former suppressed the same. Having aratio of SFs provides an opportunity to the gateway to demodulate multiple packets resulting in ahigher throughput and an increase in the number of supported nodes. On the other hand, SF12,having the highest airtime, becomes detrimental as colliding packets of same SF are not receivedcorrectly within the same channel.

The experiments in this section were performed to quantify the performance in of single physicalchannel. A study across multiple channels is required to fully understand the network capacity of asingle gateway LoRa network since LoRa concentrator chipset promises concurrent demodulationsacross channels.

4 BEYOND LORAThis section reports our studies beyond LoRa practices and provides insights through an SDR basedLoRa PHY analysis. We also report optimization and improvement opportunities to existing LoRapractices.

4.1 Parameter Optimization for Communication DistanceThis section proposes an algorithm to search for the optimized parameter setting for a givencommunication distance and a node lifetime restriction. Based on the results in Section 3.1 and3.2 which predict the communication distance and lifetime of a node respectively for any givenparameter settings, the proposed Algorithm 1 makes use of the prediction model fcoveraдe () andfl i f et ime () which are derived using transmission distance and node lifetime in years as a functionof SF and TX Pow. The models make use of experiment data which are first filtered by usingPRR threshold to remove any data points below the given PRR. Filtered data are then used in thepolynomial regression to train the models. Since the input of both models are similar, a simplesearch through the SF and TX Pow combination is used and all settings that fulfill the requirementof fcoveraдe () are then used as input to fl i f et ime ().The resulting output from fcoveraдe () and fl i f et ime () is presented in Fig. 17 where different

inter-packet delays are used for the three figures. Each setting’s coverage is plotted against itsrespective lifetime with TX Pow increasing along the x-axis. These figures are generated based onthe profile of a LoRa node transmitting 10 bytes of payload with 8 preamble symbols and code rateof CR4/8. The inter-packet delays are 15, 30, and 60 minutes respectively.According to the results in Fig. 17, SF could reach further distance but TX Pow would provide

better lifetime. As observed in the figures, lower SF with higher TX Pow provides similar coveragebut with longer lifetime. A simple optimization to the algorithm can be applied by making useof the nature of SF and TX Pow whereas SF and TX Pow increases, the coverage increases whilelifetime decreases accordingly. This nature can be incorporated into the algorithm by searching



0 1 2 3 4Coverage (km)

0

2

4

6

Life

time

(yea

rs)


(a) Inter-packet delay of 15 min-utes.


0

2

4

6

Life

time

(yea

rs)


(b) Inter-packet delay of 30 min-utes.


0

2

4

6

Life

time

(yea

rs)


(c) Inter-packet delay of 60 min-utes.

Fig. 17. Lifetime and coverage of all settings for PRR of 0.7 in NLoS environment. The skyline of all curves ineach figure is the optimal settings for the given inter-packet delay.

Algorithm 1 An optimal settings searching algorithm.Input:

Thd : Threshold for minimum required distanceThl f : Threshold for minimum required node lifetimeTcycle : Packet interval requirementfcoveraдe (): Prediction model for LoRa coveragefl i f et ime (): Prediction model for LoRa lifetime

Output:Soptimal : A set of optimal settings (s f , tx_pow)

Initialize:Soptimal ←− {}

1: for s f ∈ {SF12, ..., SF7} , tx_pow ∈ {20dBm, ..., 2dBm} do2: if fcoveraдe (s f , tx_pow) ≥ Thd then3: if fl i f et ime (s f , tx_pow,Tcycle ) ≥ Thl f then4: Soptimal ←− (s f , tx_pow) ◃ settings fulfill conditions5: end if6: else7: break ◃ skip remaining settings8: end if9: end for10: return Soptimal

through decrementing loops instead of incrementing loops. The algorithm only terminates thesearch while searching through the fcoveraдe () but not fl i f et ime (). This is mainly because the inputof fl i f et ime () is a subset of the search space and early termination would not return the skylineof Fig. 17 since many higher SF settings would have lower lifetime than the lower SF settingscounterparts. Recall that the search is done using a decrementing search where higher SF and TXPow values will precede the lower values.

We present the output of Algorithm 1 in Fig. 18 which represents the optimal settings with givenlifetime and coverage requirements. Observe that lifetime of a node is severely impacted by packetfrequencies. However, the relationship of node lifetime and packet frequencies are non-linear. This



0 0.5 1 1.5 2 2.5 3 3.5

Coverage (km)

0

1

2

3

4

5

Life

time

(yea

rs)

SF

12-2

0dB

m

SF

11-2

0dB

m

SF

10-2

0dB

m

SF

9-20

dBm

SF8-20dBmSF7-2dBm

15 min30 min60 min

Fig. 18. Setting selections for optimal coverage and lifetime. Exhausting TX Pow before SF would results in anode with better coverage and longer lifetime.

is due to the power consumed during a node’s sleep time which is small but not negligible as sleeptime increases.

The experiment results presented here provide a rule of thumb to system adopters and researchers.Since increasing SF results in lower lifetimes with higher coverage, SF should always be the lastsetting to be increased. System adopters would be able shorten the time to obtain optimal settingsusing the algorithm provided and researchers would be able to design more efficient settingadaptation mechanisms by considering the findings highlighted above.

4.2 Implication of LoRaWANThis section discusses about the LoRaWAN MAC protocol which serves as guideline for LoRa’sMedium Access Control (MAC) layer as well as upper network stacks layers and the implication ofLoRaWAN towards the performance of LoRa.LoRaWAN proposes three types of classes to cater for different application requirements. The

three classes are class A (All end nodes), B (Beacon), C (Continuous listening)[31]. Class A beingthe basic option which all LoRaWAN compliance end nodes should be able to support while class Band C being a mutually exclusive additional features on top of class A. This implies that class Cend nodes should not implement class B and vice-versa.

TX RX1 RX2TRX_DELAY1

TRX_DELAY2

RX1 RX2TRX_DELAY1

TRX_DELAY2

Beacon Beacon... ...

TX RX1 RX2 ContinuousTRX_DELAY1

TRX_DELAY2

CLASS A

CLASS B

CLASS CRX2

TBeacon

TX

Fig. 19. Classes of LoRaWAN exhibit different network utilization behaviors.



Class A devices implement a node-initiated transmission where all downlink communicationfrom the server to end nodes have to be initiated by an uplink transmission from end node to server.To enhance the reception of downlink communication, end nodes are required to schedule tworeception slots after the initial transmission. The first reception slot uses the exact settings as priortransmission while the second reception slot uses a preprogrammed SF. First and second receptionslot are initiated following a transmission after a TRX _DELAY 1 and TRX _DELAY 2 respectively. Thesereception slots only act as a preamble detection window. In any case where the packet in firstreception window is designated for the end nodes, the second reception window would not beopened. Similarly, if the first reception window duration exceedsTRX _DELAY 2, the second receptionwindow will also be aborted. LoRaWAN suggests five Tsym as timeout for each reception window.

As the name suggests, class B incorporate beacons to improve responsiveness of end devices. Endnodes are expected to obtain TBeacon from gateways and wakes up every TBeacon to synchronizeitself with gateways to open a short reception window when necessary. These reception windowwill allow gateways to transmit any command from the server to end nodes within a TBeacon .

Class C is designed for real time applications that require immediate response from end nodes.With this class, end nodes are required to open a continuous reception window using the pre-programmed SF. Instead of waiting for TRX _DELAY 1 before reception, class C devices immediatelyopen a reception window with preprogrammed SF for TRX _DELAY 1. After TRX _DELAY 1, end nodesswitches the reception settings to settings used in transmission before going back to preprogrammedsettings after TRX _DELAY 2.

LoRaWAN incorporates additional overhead into the packet to enable the Medium Access Control(MAC) features provided by classes above. These overheads include MAC headers, commands, andMessage Integrity Checks (MIC). By combining the overheads, LoRaWAN imposes 13 to 27 Bytesof overhead onto each packet transmitted by an end node.

Energy ConsumptionWith the increase in packet size due to the MAC overhead and mandatory reception windows,energy consumption of an end node would increase accordingly. An energy consumption snapshotof class A end node is illustrated in Fig. 20. Apart from the TX phase, there are two RX phasesseparated by predefined delays. Note that the MCU is not in low power mode during the delays.These MCU active time would results in degradation of an end node’s lifetime.

0 1 2 3 4

Time (s)

0

25

50

75

100

125

150

175

Pow

er (

mW

)

TX

RX1 RX2

TRX_DELAY1

TRX_DELAY2 T

MCU

SLEEPSLEEP

Fig. 20. Node energy consumption for a LoRaWAN transmission cycle. On top of the energy consumptiondue to RX, MCU has to stay awake throughout TRX _DELAY 1 and TRX _DELAY 2 which increases the MCUenergy consumption.



An experiment was conducted to explore the energy budget overhead of LoRaWAN as comparedto LoRa. The experiment measures energy consumption of a transmission cycle which includesboth active and sleep time of an end node. Energy consumption of end nodes was measured usingMonsoon power monitor and payload size was set to 10 Bytes with 0.1% duty cycle to adhere toLoRaWAN standards. For each packet, CR is set to CR4/8 and 100 packets were transmitted foreach SF and TX Pow setting combination. For LoRaWAN packets, we assume no downlink trafficfrom gateway which requires end nodes to execute both RX windows with default RX timeoutof five symbol time. Energy consumption are first grouped according to the settings followed bygrouping of energy consumption according to component states (EMCU _OF F , EMCU _ON , ER_OF F ,ER_TX , ER_RX 1, and ER_RX 2). After grouping, each group are then averaged to obtain the energyconsumption of each states for each setting. Experiment results are presented in Fig. 21.

SF7 SF8 SF9 SF10 SF11 SF120

0.2

0.4

0.6

0.8

1

1.2

Ene

rgy

(J)

EMCU_OFF

EMCU_ON

ER_OFF

ER_TX

ER_RX1

ER_RX2

(a) LoRa.

SF7 SF8 SF9 SF10 SF11 SF120

0.2

0.4

0.6

0.8

1

1.2

Ene

rgy

(J)

EMCU_OFF

EMCU_ON

ER_OFF

ER_TX

ER_RX1

ER_RX2

(b) LoRaWAN.

Fig. 21. Energy budget required by a transmission cycle for packet with payload of 10 Bytes and 0.1% dutycycle with varying SF and TX Pow. LoRaWAN energy budget for EMCU _OF F and EMCU _ON is significantlyhigher than the LoRa counterparts.

Results show that LoRaWAN imposes 149% to 313% overhead on energy consumption as comparedto LoRa with average of 198% times increment across all settings. This increment is due to theincrease in MCU active time while waiting for RX windows mandatory in LoRaWAN standard.However, note that with acknowledgment (ACK) packets from gateway, such overhead couldbe reduced as end nodes are expected to not start the second RX window if ACK or downlinkpackets are received in the first RX window. On the other hand, any long downlink packet couldsignificantly increase the energy overhead.

Multiple AccessOn top of the overhead in energy consumption, a MAC protocol also imposes limitations on packettransmission. This limitation affects several network parameters which includes network density,throughput, and goodput. Since LoRa networks are expected to form a single hop network withoutintermediate end nodes as relay, LoRa network density directly equates to number of end nodessupported by a gateway. This value could vary depending on the PRR requirement of said network.Similarly, throughput and goodput which refers to the end-to-end number of bytes received andnumber of effective bytes received respectively also varies depending on PRR but are unaffected bymulti-hop.To investigate the effect of LoRaWAN imposed limitation on parameters mentioned above,

experiments were conducted using a SX1301 gateway concentrator with 50 end nodes equippedwith SX1276 chipset. Similar to the experiment in Section 3.3, 50 nodes were used to emulates



multiple nodes bymanipulating the inter-packet delay.We first assume a standard transmission withinter-packet delay of 15 minutes regardless of SF, then by halving the inter-packet delay, numberof active nodes increase to 100. Using said method, number of emulated nodes transmitting in 15minutes could reach up to 500,000 nodes. Next Eq-9 were used to vary the probability of a node inchoosing a particular SF for transmission to simulate outdoor, semi-indoor, and indoor environment.Furthermore, in order to ensure no interruption due to ACK queuing and self-interference fromACK packets, downlink traffic was disabled. However, by disabling downlink traffic, end nodeswere expected to activate both RX windows. TX Pow of the nodes are set to the maximum value of20dBm while CR are set to CR4/8. Payload size of each packets are set to 10 Bytes to aid comparingwith results in Section 3.3.

(a) Number of nodes and throughput supportedby each PRR requirement.

(b) Number of nodes and goodput supported byeach PRR requirement.

Fig. 22. Comparison of throughput and goodput of a single gateway with LoRaWAN and supported numberof nodes for each PRR requirement. Similar to results in Section 3.3, higher number of SF7 packets increasesthroughput, goodput, and number of supported nodes.

Experiment results in Fig. 22 shows a significant advantage of more SF7 packets in the network ascompared to equal chances which signifies indoor and semi-indoor respectively. Moreover, havingmany SF12 in the network could cause congestion and lower network performance. Comparingthroughput and goodput of LoRaWAN system reveals that LoRaWAN packet overhead accounts to>64% of the traffic however, this overhead would reduce as the payload size increases.

LoRaWAN which is the MAC protocol of LoRa serves to organize the channel access behaviourof nodes within the network. This mechanism should provide certain degree of improvement overLoRa’s disorganized behaviours. A comparison between results of LoRa and LoRaWAN which arepresented in Fig. 23 indicates that although in most cases LoRaWAN provides a higher throughput,the header overhead of LoRaWAN takes up most of the throughput. When observing goodputwhich is throughput without the overhead, benefit reported in throughput vanishes leaving behinda much lower goodput compared to aloha LoRa. The high throughput reported was mainly dueto the mandatory RX windows. These RX windows provide a chance for other nodes to makeuse of a clear channel to transmit packets. However, the same would not be true when downlinktraffic is not disabled. Downlink traffics, especially ACK packets could block a significant amountof uplink traffics since in order to transmit an ACK packet, gateway has to switch one of its tworadio chains from RX mode to TX mode. This change in modes would disable all uplink trafficstransmitting in the channels served by that particular radio chain within the period of time theACK is transmitted. Moreover, a tight scheduling is essential to ensure minimal delay in de-queuingreceived packets and queuing ACKs. In order to offload the burden of a gateway in providing ACKs,all ACKs requests would be served by backend servers.



(a) Difference between LoRa and LoRaWAN in sup-ported number of nodes and throughput for eachPRR requirement.

(b) Difference between LoRa and LoRaWAN in sup-ported number of nodes and goodput for each PRRrequirement.

Fig. 23. Comparison of the difference in throughput and goodput of a single gateway between LoRa andLoRaWAN with supported number of nodes for each PRR requirement. Difference is computed as δx =LoRaWAN x

prr − LoRaxprr where x refers to #nodes, throughput, or goodput. Hence, positive and negative δx

are represented in the figure with red and gray bars respectively. LoRaWAN packet overhead diminishes thebenefit it brings.

With the unique capability of LoRa gateway which allows concurrent demodulation, an efficientMAC protocol is required which takes into consideration of providing high packet reception ratewith minimal ACKs and retransmission.

4.3 LoRa Medium AccessLoRa’s physical layer is aloha in nature and incorporates no collision avoidance mechanisms.However, lost transmissions are undesirable within sensor networks and incur unbearable costsupon energy restricted nodes. In LoRa, only packets of same SF within a common channel aresusceptible to disruptive collisions and the probability of such collisions increases with increasingnetwork density. Traditional RSSI based methods used for collision avoidance are ineffectiveon LoRa for two reasons. 1) Due to inherent properties from CSS modulation, successful LoRatransmissions may traverse below noise floor. 2) RSSI itself offers no capability to differentiatebetween single channel concurrent transmissions. In this section, we focus on the performance ofa LoRa concentrator in receiving orthogonal and non-orthogonal transmissions within single andmultiple channels through a series of experiments. Based on their observations, we then propose apotential collision avoidance mechanism for LoRa, named CSMA-CAD, which improves PRR underdense network conditions.

Concurrent Reception Capacity of the LoRa ConcentratorLoRa’s concurrency is highly reliant on orthogonal transmissions. Therefore, prior to discussingconcurrency, we first illustrate orthogonal LoRa transmissions in the physical layer. To do this, wetransmit three concurrent packets using distinct Spreading Factors SF7, SF9, SF11 within a singlechannel. At the same time, to better illustrate the chirps, we also concurrently transmit the samedata on three different channels. We then capture this concurrent transmission using the SDR setupdescribed in Section 2. Fig. 24 presents a spectrogram of these LoRa transmissions. All channelswere of 125KHz of BW. While CH2, CH3 and CH4 have only a single transmitting node, CH1 hasthree nodes set on three orthogonal SFs. A transmission of this sort is seen by the concentrator asa concurrent transmission of six packets via four channels.In addition, Fig.24 also leads to validate the theoretical relationship between Tsym and SF as

predicted by Eq-1. According to Eq-1, the symbol durations for SF7, SF9 and SF11 should be ~1ms ,


Known and Unknown Facts of LoRa 19:25C

H1

CH

2C

H3

CH

4

Frequency

(K

Hz)

SF11

SF11

SF9

SF7

SF7SF9

0ms 10ms 20ms 30ms 40ms

Fig. 24. Concurrent transmission of LoRa.

~4ms , and ~16ms respectively which corresponds to doubling of Tsym per each incrementing SF.Therefore, Tsym for SF11 should be 16 times Tsym of SF7. The visualized symbols on CH4 clearlyattest to this statement. For example, during a single SF11 chirp, there exists 16 SF7 and 4 SF9chirps. However, this should not lead to the conclusion that for a given payload, LoRa modulationdoubles the dwell time for incrementing SFs. This is due to zero padding causing different numberof data bits per each SF defined in the LoRa modulation resulting in a different number of totalsymbols for the same payload.Having highlighted the advantages of orthogonal channels in LoRa, it can be observed that

successful concurrent transmissions need not be limited to physically separated channels. Next, weestimate the extent of this concurrency of a standard IC880A LoRa concentrator.

Table 6. PRRs under varying concurrent conditions. With separation of either SF or CH between packets,LoRa gateway is able to demodulate collided packets efficiently.

#Concurrent Tx 2 3 4 5 6 7 8Multiple CH, Multiple SF 100% 100% 100% 100% 100% 100% 99%Multiple CH, Fixed-SF 100% 100% 100% 100% 100% 100% 100%Single CH, Multiple SF 100% 100% 100% 100% 100% – –Single CH, Fixed-SF 48% 32% 24% 19% 16% 13% 12%

To do this, we perform four experiments. In the first, we tie eight nodes to eight distinct chan-nels. We program each node to continuously transmit 100 packets on a randomly selected SF. Inall, to ensure continuous concurrency throughout the experiment, each transmission was timesynchronized across all nodes. This synchronization was required to provide sufficient time to anode that opted to transmit at a higher SF (consuming a higher airtime) to finish the transmission.Since this experiment ensured eight packets in air at any given time, it can be used to validate



Semtech’s claim that the SX1301 LoRa concentrator can decode and demodulate eight packets atany given time. In the second, the same experiment was repeated by fixing all transmissions SF toSF7 on eight different channels. Results from both experiments depicted in the first two rows ofTable 6 attest to the aforementioned claim by Semtech.

In the third, we test the concentrators capacity to receive orthogonal transmissions. We tiesix nodes to six distinct SFs and program them to continuously transmit 100 packets on thesame channel. The limitation on the number of nodes was due to the limitation on available SFs.Again, time synchronization between nodes ensured continuous concurrency throughout theexperiment. As expected, the results indicated 100% PRR. Similarly, in the fourth, we transmit non-orthogonal packets within same channel. Our results confirm the disruptiveness of non-orthogonaltransmissions. The results of third and fourth experiments are depicted in third and fourth rows inTable 6.

A Potential CSMA Mechanism for LoRa

(a) The CSMA-CAD mechanismillustrated as a state diagram.

2 4 6 8 10Inter Packet Delay (s)

0

0.2

0.4

0.6

0.8

1

PR

R

CADNo CAD

(b) Comparison of CSMA-CAD and No-CSMAmechanisms for varyingTpkt . PL=6B,preamble length=9 and #Nodes=50.

Fig. 25. CSMA-CAD states and experiment results.

An ideal mechanism to avoid potential collisions in LoRa is to sense the channel for non-orthogonal SFs prior to transmission. If a LoRa transceiver is able to detect such transmissions,it forms the foundation to formulate an ideal CSMA mechanism for LoRa. Unfortunately, theSX1276 series LoRa transceivers limit such sensing to the preamble and calls it the Carrier ActivityDetection (CAD)-Mode [45]. Despite the limited functionality of CAD-Mode, we leverage the CAD-Mode to form CSMA-CAD and investigate the effectiveness of utilizing it as a potential collisionavoidance mechanism. Although longer preambles could significantly increase the effectiveness ofCSMA-CAD mechanism, without further changes on the concentrator, we confine our experimentsto 8 preamble symbols mandated in the LoRaWAN standard. CAD-Mode requires a duration oftwo symbols to complete. During the CAD process, the radio first performs a receive operationfor a duration of a single symbol followed by a computation duration of another symbol. ThisCSMA-CAD process for an end node is described by Fig. 25a.We then designed an experiment to test whether incorporating CAD-Mode is beneficial in a

real-life environment. To do this, 50 nodes were fixed to a single channel and were programmed totransmit 100 packets of 10 Bytes payload. Each node transmitted 100 packets randomly within aninter-packet delay of Tpkt . A random start delay ensured the distribution of transmissions across



0

20

40

60

80

100

120

140

160

180

200

0 1 2

Power (m

W)

TTX TTXTMCU

TX TX

CAD‐Mode

Fig. 26. No-CSMA vs CSMA-CAD energy comparison.

Tpkt which allowed the control of network density. We computed PRR for each experiment andrepeated the same forTpkt ranging from 2 to 10 seconds. A payload of 4 Bytes were chosen and thelongest airtime corresponding to the airtime of SF12 was ensured to be less than Tpkt .As indicated by the experiment results in Fig. 25b the CAD-Mode enabled significant improve-

ments to PRR under high network density conditions, i.e. when Tpkt is low. Further, we alsodetermined the highest possible energy overhead (SF12 with longest Tsym ) due to CAD-Mode to beat only 1.6963mJ. We present below a comparison of power profiles for both modes in Fig. 26 trans-mitting a payload of 4 Bytes under SF12 at 5dBm. The peak indicated on the second transmissionrepresents the power overhead of CSMA-CAD. I.e. CSMA-CAD sensing and applying a different SF.As ideal CSMA protocol for LoRa should be able to sense the on-going Spreading Factor at anypoint of payload, not just within the preamble. Nevertheless, based on our experiment results, itcan be seen that performance improvements are still possible with the CSMA-CAD mechanism.

We then extend the boundaries of this experiment by means of a simulation in an effort to test theeffect of varying number of nodes, Tpkt and preamble length on the PRR. The simulation considersan erasure channel and compares the three CSMA mechanisms; 1. No-CSMA 2. CSMA-CAD 3.Ideal-CSMA. Despite the obvious advantages that come through an Ideal-CSMAmechanism, CSMA-CAD still provides significant performance improvement under dense network conditions similarto experiment results.

It can be seen from the results that the Ideal-CSMA mode always achieves a full packet receptionratio except in Fig. 27a where the number of nodes reached up to 500. This is understandable asmany nodes perform symbol sensing, it is possible that a few performed synchronous symboldetection and transmitted simultaneously resulting in collisions. Despite the existence of this effectwithin a simulated environment, the same phenomena can take place within real-life environmentscausing potential decrements in PRR. Lastly, Fig. 27c depicts that the Ideal-CSMA mechanism isnot significantly affected by the number of preamble symbols. However, the impact is significantfor a CSMA-CAD network which achieves a doubled performance in PRR with only 4 additionalpreamble symbols. At the same time, as expected, the No-CSMA mechanism reflects an inverseperformance affected by the increased preamble symbols.Semtech has also began to answer the need for a better CSMA mechanism in LoRa chipsets

by announcing a new chipset which allows CSMA mechanism beyond CAD [57]. With improvedCSMA mechanisms, researchers would be able to design better MAC protocols which would reducethe number of possible collisions while enhancing network quality.

4.4 Revisiting LoRa PHYUnderstanding LoRa PHY would open a path to enabling many applications such as passivelocalization [53], cooperative transmission [14, 15], co-existing and complementing MAC protocol



(a) PRR vs #Nodes where preamble=8, pay-load=6B and Inter-packet Delay fixed at 15s.

(b) PRR vs Inter-packet Delay where pream-ble=8, payload=6B and number of Nodes fixedto 100.

(c) PRR vs Preamble Length where payload=6B,Inter-packet Delay=15s and number of Nodesfixed to 100.

Fig. 27. Simulation of a LoRa network under No-CSMA, CSMA-CAD and Ideal-CSMA mechanisms. AlthoughIdeal-CSMAmechanism shows promising results, hidden terminal problem and synchronous symbil detectionwould still cause packet collisions.

[19], etc. To understand LoRa PHY, one has to start from the encoding and modulation process of apacket.

For a given payload, LoRa PHY integrates several encoding mechanisms to improve the over theair resiliency prior to modulation. This encoding process involves Forward Error Correction (FEC),Interleaving, Whitening and Gray Coding in order[48]. The process of modulation is responsiblefor integrating data bits into "base chirps" also known as "linear chirps" explained duly. We provideinsights to this process under this section.

LoRa Modulation and DemodulationThe complex baseband representation of a single frequency (pure tone) can be expressed as followswhere f0 refers to the frequency of tone and ϕ represents a possible phase shift.

ftone (t) = e j(2π f0t+ϕ) (10)In contrast, a linear chirp or an "unmodulated" chirp sweeps through a given channel bandwidth

BW from fstar t to fend by continuously changing the frequency at ratem. Eq-11 expresses such achirp where f0 represents the starting frequency at t = 0 hence,mt+ f0 represents the instantaneousfrequency at any given time t [32].

fchirp (t) = e j(2π (mt+f0)t+ϕ) (11)A rf symbol in LoRa is a single chirp that sweeps the bandwidth BW. It can be seen that in order

to sweep the bandwidth BW in exactly Tsym duration, the rate of change of frequencym of the



chirp should bem = BWTsym

. The symbol duration Tsym is defined as Tsym = 2SFBW in Semtech’s LoRa

patent [50]. This implies that as long as the BW is fixed, a higher SF requires a longerTsym duration.A LoRa packet consists of a combination of such unmodulated as well as modulated chirps. Theunmodulated formulate the preamble and the Start Frame Delimiter (SFD) whereas the modulatedform header, payload and CRC.

In order to modulate data bits into a base chirp, its instantaneous frequency is changed exactlyonce. For any given SF, a modulated chirp will consist of a single shift out of possible 2SF − 1.This implies that depending on the SF, a single modulated LoRa chirp carries SF number of bits ofinformation. In other words, an integer in the range of 0 to 2SF − 1 is represented using 2SF − 1possible frequency shifts. During the process of demodulation, each frequency shift in the datachirp is recognized as a separate FFT bin which relates back to the original integer. In a moretechnical perspective, the process of modulation can be visualised in the time domain as follows.

After performing FEC, Interleaving, Whitening on the payload, to modulate the resulting encodeddata in to data chirps, the data is divided into SF sized bit chunks. Each chunk, of SF bits, representsan integer i within the range 0 to 2SF − 1. i is then modulated to a base chirp by introducing a timeshift of t̂ = Gray−1(i) T

2SF [48] to the signal presented in equation Eq-11 whereGray−1 stands forthe reverse of Gray operation in [18].

Fig. 28. LoRa demodulation to FFT bins.

Later, at the receiver, each received chirp is multiplied by a locally generated conjugate base chirpof same rate of change of frequencym, known as de-chirping. Multiplying two in-phase base chirpsof samem but of opposite signs results in an argmax at the bin corresponding to 0Hz in a FFT thatis of 2SF bins wide, due to destructive superposition of these two waveforms [33]. However whena received data chirp undergoes de-chirping, the argmax of the FFT for each received chirp willoccur at one of 2SF bins. This bin can be directly correlated to the originally modulated integer iwhich is in the range of 0 to 2SF − 1.



However, it should be mentioned that for a given difference in phase ϕ between the receivedand the locally generated chirp, FFT bins upon de-chirping operation may contain an offset. Thisproblem can be rectified by consulting the information from preamble and SFD bits mandated tothe front of every LoRa payload.

To further illustrate the process of demodulation, Fig. 28a closely represents a received SF8 LoRasignal of 15 payload symbols. Fig. 28b presents the same signal upon de-chirping while Fig. 28cpresents the FFT bins that correspond to i . Since the signal is SF8 modulated, each data chirp canonly take a value between 0-255 and the FFT bins are also limited to the same range. Therefore, theFFT width is set to 28 and each bin in Fig. 28c represents the data previously modulated into eachdata chirp after the preamble and SFD in Fig. 28a.

Doppler EffectWhen a transmitter or a receiver move relative to each other, the apparent frequency shift on thereceived signal is known as the Doppler’s effect. Depending on the extent of this effect, the signalis either received or discarded. CSS on the other hand is widely known to be immune to Dopplereffect [56, 65]. LoRa, being a close variant of CSS inherits a degree of resilience from it. Under thissection, we conduct a series of experiments to study possible effects caused due to Doppler shift onLoRa’s modulation under practical circumstances.

This apparent change in frequency of a LoRa chirp due to Doppeler’s shift can be integrated toLoRa’s modulation equation as follows,

fchirp (t) = e j(2π (mt 2+f0t+∆fd t )+ϕ) (12)

where ∆fd is the change frequency caused due to Doppler effect. Under general use cases whereend nodes move at a velocity ven relative to a stationary LoRa gateway, this ∆fd approximatesto ∆fd =

venC fchirp where fchirp represents the instantaneous chirp frequency. In the case of CSS,

this ∆fd causes the autocorrelation peak on the receiver side to shift in time and this time shift isapproximated by ∆fd

m [65]. The same approach can be used to determine the most susceptible SFs inLoRa modulation under Doppler effect. It can be seen for smaller chirp ratesm, the time shift ishigher. This concludes that SF12 is most effected while SF7 is the least.

In order to quantify the effect of Doppler shift on LoRa chirps we set-up a mobile LoRa gatewayand an SDR setup by the roadside and attach a node to a car used as moving transmitter. We drive thecar at varying speeds ranging from 50kmph, 60kmph, 70kmph and 80kmph on a highway. Extremecare was taken at all times to maintain persistent speed of the car and line of sight to the gatewayduring transmissions. Each transmission with 50 bytes were performed on SF12 with a packetduration of 2.35 seconds to ensure sufficient time for recording a packet with both approaching andpassing states. Each round maintains a fixed speed with the vehicle approaching the gateway, in-between and lastly passing the gateway. More specifically, each transmission consisted a sequencenumber. A gateway was placed by the highway ensuring LOS with the car while a second receivedrecords the transmissions from inside the vehicle without Doppler’s effects.Our results indicate strong packet reception (>85%) at gateway placed by the highway for all

tested speeds and Table 7 summarize the results of the experiment for all speeds.Further, we also compare recorded transmissions from within the car with the transmissions

recorded using the receiver placed by the highway using the SDR setup seeking for observableDoppler effects. We present a comparison of two snapshots of LoRa chirps transmitted at 80kmph inFig. 29. Visually, both recordings offered no observable difference and similar results were observedfor other speeds. On contrary to the findings in [39] who reported a significant influence on PRR



Table 7. Experiment results for doppler test with varying speeds.

Speed(kmph) #RX/TX PacketsApproaching Intermediate Leaving

50 1/1 1/1 1/260 3/4 1/1 3/370 3/3 1/1 3/380 2/2 1/1 1/2

even at speed 8kmph, our results show no observable influence on PRR nor spectogram recordseven at high speed of 80kmph.

Chirps recorded from inside the vehicle

Chirps recorded at the gateway

Freq

uenc

y (K

Hz)

Time

Time

Fig. 29. A LoRa transmission recorded from inside the vehicle (top) and at the gateway (bottom) for a payloadof 50 Bytes at SF12. The vehicle was driving pass the gateway at the speed of 80kmph. No significant differenceon the chirps can be observed from comparing the two spectograms.

Although LoRa technology is robust to a high degree against Doppler effect, it may still bepossible to capture frequency deviations in the chirps due to Doppler effects and predict the speedof moving end nodes [22, 38].

5 RELATEDWORKS AND CONCLUSIONIn this paper, we have conducted experiments to verify common claims made on LoRa and derivedmodels in accordance with said promises.A significant amount of work has been focused on differences between LoRa and similar

technologies[12, 25, 28, 36, 55, 63] including batteryless operation of LoRa enabled devices[40, 60].Few others have also focused on surveying various aspects of the LoRa network stacks [26, 44].Our work differs from those as it concludes a series of experiments that help derive answers toquestions regarding LoRa’s promises and also beyond.

In verifying LoRa’s communication distance, this paper answers important questions such as "howfar can a LoRa end node transmit?" or "what parameters should a node use for optimal coverage?"etc. Through multiple experiments under a variety of environments, the results presented inthis paper show that LoRa transmissions are capable of transmitting beyond 10km under LoSenvironments. However, with the presence of obstacles blocking LoS, LoRa’s coverage declinesdrastically. Moreover, inmetropolitan city environments among high-rise buildings, LoRa’s coveragedrops to 3km. This paper also quantifies the importance of SF over other parameters and providesboth researchers and system adopters a rule of thumb to always select the lowest SF wheneverpossible. The results in this paper agrees with the those presented in [4, 29, 30, 37, 62] for coveragein open environments and [20, 42] for coverage in indoor environments. In contrast to [13] where



signal propagation is modeled for different environments, this paper tries to model the effect ofdifferent parameters on coverage under varying environments.

Another important question this paper tries to answer is "how do modulation parameters affectlifetime of nodes?". In doing so, this paper proposes a model to quantify the lifetime of a node bypredicting the energy consumption during a transmission. The proposed model was derived fromreal measurement data across multiple chipsets. With it, the paper concludes that a node could onlysustain its operation to the order of years with careful parameter selection, extremely low dutycycle setting and meticulous selection of MCUs. Moreover, LoRaWAN imposes on average 198% ofoverhead to the energy budget of a node by increasing both payload size and node active duration.A similar study was done in [9] which proposed a model to compute the energy expended by anode during a transmission cycle however, our work found that the expended energy is non-linearand can be modeled. Moreover, Petajajarvi et. al [43] highlights the power consumption of a nodeas a whole. Our work takes this step further by dissecting a transmission cycle along with itssub-components responsible for the lifetime of a node.A fundamental question raised prior to deploying a LoRa network would be "how many nodes

can a gateway support?". This paper performed multiple experiments to quantify the node capacityof a gateway through 1) quantifying the number of supported nodes for a single SF network on asingle channel, 2) quantifying the number of supported nodes for a single channel with multipleSFs, and 3) exploring the overheads and gains of LoRaWAN. Counter-intuitively a channel withonly SF7 supports a lower number of nodes as compared to a channel with only SF12 due to highnumber of packet collisions caused by high number of inter-packet interference from the low dutycycle of SF7. On the other hand, a mixed SFs network performs better if more SF7 packets arepresent in the network. This is due to the concurrent demodulation capability of the gateway whichdemodulates collided packets in the orthogonal space. While a single SF channel would be full ofnon-orthogonal packet collisions, a mixed SF channel would innately leverage orthogonality of SFsto avoid collisions along with capture effect [21] to increase both throughput and the number ofsupported nodes. Results of the experiments indicate that a single channel gateway with multipleSFs could support >14,000 nodes. While simply extending the capability of a single channel tomultiple channels would not be logical however, a single channel with six SFs only exploits sixout of eight concurrent demodulators designed in the concentrator chipset. Additionally, as thestandard MAC protocol for LoRa, LoRaWAN overhead does not provide equivalent returns. Ontop of the high energy overhead due to encryption and RX delays, significantly large LoRaWANoverhead due to header information diminishes the throughput gain it brings. Albeit similar to[2, 5, 6, 8] who focuses on the network capacity of LoRaWAN, this paper mainly explores thenumber of nodes supported by LoRa itself and hopes to serve as a baseline for future studies.Looking further into the combined experiment results, this paper answers the question every

system adopter would like to know, "what is the optimal setting for a node?". To answer thisquestion, the paper proposes a simple search algorithm to find the optimal setting and also providesrule that suggest that TX Pow is a more efficient parameter in terms of end node lifetime whileSF is a more efficient parameter in terms of coverage. While there exists multiple SF and TX Powsettings to reach a desired distance, settings with lower SF and higher TX Pow should always bepreferred which reduces the search space by 79%. With this significant reduction in the searchspace, system adopters would be able to survey deployment areas more efficiently while researcherswould be able to design more efficient rate adaptation algorithms. Reynders et. al [47] proposea parameter selection algorithm by evaluating individual path loss values of each node on eachchannel assessed by the gateway which then assign the selected parameter to the end node throughdownlink transmissions. Results in this paper could compliment Reynders et. al algorithm bylimiting the search space.



Although a LoRa concentrator’s capability to demodulate orthogonally colliding packets is incontrast to other LPWAN technologies, a LoRa gateway could still experience collisions from packetswithin non-orthogonal data rates. Studies like CHOIR [15] has proposed a method to separatesuch collisions by leveraging imperfections present within an end node’s hardware. However,collisions could be averted by leveraging a CSMA mechanism prior to transmission. CAD mode inthe SX127X chipset is capable on sensing a potentially colliding packet during the preamble. Thispaper exploits this mode by implementing a simple CSMA mechanism and discovers that paying asmall price in energy budget (< 1.70mJ ) for CSMA-CAD provides up to 20% PRR improvement.A simulation was also conducted to investigate the possible gain of an ideal CSMA mechanism.The simulation predicts that an ideal CSMA mechanism can increase the throughput up to 56×compared to CSMA-CAD.

In summary, this paper has answered common questions system adopters and researchers haveabout LoRa and verifies the promises Semtech has made in terms of transmission distance, end nodelifetime, and node capacity. It further provides a deeper understanding on the effect of modulationparameters on the aforementioned promises which could aid system designers in designing a betterMAC protocols for LoRa. Lastly, a simple experiment to test the resiliency of LoRa against theDoppler effect also resulted encouraging results about LoRa modulation.

REFERENCES[1] LTE Release 13. 2018. Release 13. Retrieved Aug 8, 2018 from http://www.3gpp.org/release-13[2] Ferran Adelantado, Xavier Vilajosana, Pere Tuset-Peiro, Borja Martinez, Joan Melia-Segui, and Thomas Watteyne.

2017. Understanding the limits of LoRaWAN. IEEE Communications magazine 55, 9 (2017), 34–40.[3] Wi-Fi Alliance. 2018. Wi-Fi HaLow | Wi-Fi Alliance. Retrieved Aug 8, 2018 from https://www.wi-fi.org/discover-wi-fi/

wi-fi-halow[4] Aloÿs Augustin, Jiazi Yi, Thomas Clausen, and William Mark Townsley. 2016. A study of LoRa: Long range & low

power networks for the internet of things. Sensors 16, 9 (2016), 1466.[5] Eyuel D Ayele, Chiel Hakkenberg, Jan Pieter Meijers, Kyle Zhang, Nirvana Meratnia, and Paul JM Havinga. 2017.

Performance analysis of lora radio for an indoor iot applications. In Internet of Things for the Global Community(IoTGC), 2017 International Conference on. IEEE.

[6] Dmitry Bankov, Evgeny Khorov, and Andrey Lyakhov. 2016. On the limits of LoRaWAN channel access. In Proceedingsof the 2016 International Conference on Engineering and Telecommunication (EnT), Moscow, Russia. 29–30.

[7] Steve Battle and Benedict Gaster. 2017. LoRaWAN Bristol. In Proceedings of the 21st International Database Engineering& Applications Symposium. ACM, 287–290.

[8] Norbert Blenn and Fernando Kuipers. 2017. LoRaWAN in the wild: Measurements from the things network. arXivpreprint arXiv:1706.03086 (2017).

[9] Taoufik Bouguera, Jean-François Diouris, Jean-Jacques Chaillout, Randa Jaouadi, and Guillaume Andrieux. 2018.Energy Consumption Model for Sensor Nodes Based on LoRa and LoRaWAN. Sensors (Basel, Switzerland) 18, 7 (2018).

[10] Michael Coracin and Matthieu Leurent. 2017. GitHub - Lora-net/lora_gateway: Driver/HAL to build a gateway using aconcentrator board based on Semtech SX1301 multi-channel modem and SX1257/SX1255 RF transceivers. RetrievedAug 8, 2018 from https://github.com/Lora-net/lora_gateway

[11] CRC Computation 2013. Implementing Data Whitening and CRC Calculation in Software on SX12XX Devices. Rev. 1.[12] Jonathan de Carvalho Silva, Joel JPC Rodrigues, Antonio M Alberti, Petar Solic, and Andre LL Aquino. 2017. LoRaWAN-

A low power WAN protocol for Internet of Things: A review and opportunities. In Computer and Energy Science(SpliTech), 2017 2nd International Multidisciplinary Conference on. IEEE, 1–6.

[13] Dalibor Dobrilović, MilanMalić, DušanMalić, and Srđan Sladojević. 2017. Analyses and optimization of Lee propagationmodel for LoRa 868 MHz network deployments in urban areas. Journal of Engineering Management and Competitiveness(JEMC) 7, 1 (2017), 55–62.

[14] Wan Du, Jansen Christian Liando, Huanle Zhang, and Mo Li. 2017. Pando: Fountain-enabled fast data disseminationwith constructive interference. IEEE/ACM Transactions on Networking 25, 2 (2017), 820–833.

[15] Rashad Eletreby, Diana Zhang, Swarun Kumar, and Osman Yağan. 2017. Empowering Low-Power Wide Area Networksin Urban Settings. In Proceedings of the Conference of the ACM Special Interest Group on Data Communication. ACM,309–321.


http://www.3gpp.org/release-13

https://www.wi-fi.org/discover-wi-fi/wi-fi-halow

https://www.wi-fi.org/discover-wi-fi/wi-fi-halow

https://github.com/Lora-net/lora_gateway


[16] Eric Gourlaouen, Johan Stokking, and Hylke Visser. 2017. GitHub - TheThingsNetwork/packet_forwarder: Packetforwarder for Linux based gateways. Retrieved Aug 8, 2018 from https://github.com/TheThingsNetwork/packet_forwarder

[17] Claire Goursaud and Jean-Marie Gorce. 2015. Dedicated networks for IoT : PHY / MAC state of the art and challenges.EAI endorsed transactions on Internet of Things (Oct. 2015). https://doi.org/10.4108/eai.26-10-2015.150597

[18] Frank Gray. 1946. Pulse code communication. US Patent US2632058A.[19] C. Gu, R. Tan, X. Lou, and D. Niyato. 2018. One-Hop Out-of-Band Control Planes for Low-Power Multi-Hop Wireless

Networks. In IEEE INFOCOM 2018 - IEEE Conference on Computer Communications. 1187–1195. https://doi.org/10.1109/INFOCOM.2018.8486301

[20] Jetmir Haxhibeqiri, Abdulkadir Karaagac, Floris Van den Abeele, Wout Joseph, Ingrid Moerman, and Jeroen Hoebeke.2017. LoRa indoor coverage and performance in an industrial environment: Case study. In Emerging Technologies andFactory Automation (ETFA), 2017 22nd IEEE International Conference on. IEEE, 1–8.

[21] Jetmir Haxhibeqiri, Floris Van den Abeele, Ingrid Moerman, and Jeroen Hoebeke. 2017. Lora scalability: A simulationmodel based on interference measurements. Sensors 17, 6 (2017), 1193.

[22] C. Huang, T. Teng, and D. Sun. 2016. Low-speed moving target detection of single frame image based on Doppler shiftestimation. In 2016 IEEE/OES China Ocean Acoustics (COA). 1–4. https://doi.org/10.1109/COA.2016.7535703

[23] IC880a 2015. WiMOD iC880A. Ver. 0.5.[24] Inair9b [n. d.]. inAir9 Datasheet and Schematics.[25] Gabor Kecskemeti, Giuliano Casale, Devki Nandan Jha, Justin Lyon, and Rajiv Ranjan. 2017. Modelling and simulation

challenges in internet of things. IEEE cloud computing 4, 1 (2017), 62–69.[26] Oratile Khutsoane, Bassey Isong, and Adnan M Abu-Mahfouz. 2017. IoT devices and applications based on

LoRa/LoRaWAN. In Industrial Electronics Society, IECON 2017-43rd Annual Conference of the IEEE. IEEE, 6107–6112.[27] Alexandru Lavric and Adrian Ioan Petrariu. 2018. LoRaWAN communication protocol: The new era of IoT. In 2018

International Conference on Development and Application Systems (DAS). IEEE.[28] Alexandru Lavric and Valentin Popa. 2017. A LoRaWAN: Long range wide area networks study. In Electromechanical

and Power Systems (SIELMEN), 2017 International Conference on. IEEE, 417–420.[29] Lingling Li, Jiuchun Ren, and Qian Zhu. 2017. On the application of LoRa LPWAN technology in Sailing Monitoring

System. In Wireless On-demand Network Systems and Services (WONS), 2017 13th Annual Conference on. IEEE, 77–80.[30] Alexander Liljegren and Robin Franksson. 2018. Measuring a LoRa Network: Performance, Possibilities and Limitations.[31] LoRaWAN 2017. LoRaWAN™ 1.1 Specification. Ver. 1.1.[32] Steve Mann and Simon Haykin. 1991. The chirplet transform: A generalization of Gabor’s logon transform. In Vision

Interface, Vol. 91. 205–212.[33] Knight Mathew and Seeber Balint. 2016. Decoding LoRa:Realizing a Modern LPWAN with SDR. In GNU Radio

Conference. GNU Radio.[34] Modulation 2015. LoRa™ Modulation Basics. Rev. 2.[35] Pierre Neumann, Julien Montavont, and Thomas Noël. 2016. Indoor deployment of low-power wide area networks

(LPWAN): A LoRaWAN case study. In Wireless and Mobile Computing, Networking and Communications (WiMob), 2016IEEE 12th International Conference on. IEEE, 1–8.

[36] Umber Noreen, Ahcène Bounceur, and Laurent Clavier. 2017. A study of LoRa low power and wide area networktechnology. In Advanced Technologies for Signal and Image Processing (ATSIP), 2017 International Conference on. IEEE,1–6.

[37] Rúben Oliveira, Lucas Guardalben, and Susana Sargento. 2017. Long range communications in urban and ruralenvironments. In 2017 IEEE Symposium on Computers and Communications (ISCC). IEEE, 810–817.

[38] I. Omelchuk, I. Prokopenko, and I. Chyrka. 2016. Multichannel target speed estimation by a colocated Doppler-pulse MIMO radar. In 2016 International Conference Radio Electronics Info Communications (UkrMiCo). 1–5. https://doi.org/10.1109/UkrMiCo.2016.7739601

[39] Dhaval Patel and Myounggyu Won. 2017. Experimental Study on Low Power Wide Area Networks (LPWAN) forMobile Internet of Things. In Vehicular Technology Conference (VTC Spring), 2017 IEEE 85th. IEEE, 1–5.

[40] Yao Peng, Longfei Shangguan, Yue Hu, Yujie Qian, Xianshang Lin, Xiaojiang Chen, Dingyi Fang, and Kyle Jamieson.2018. PLoRa: a passive long-range data network from ambient LoRa transmissions. In Proceedings of the 2018 Conferenceof the ACM Special Interest Group on Data Communication. ACM, 147–160.

[41] Juha Petäjäjärvi, Konstantin Mikhaylov, Marko Pettissalo, Janne Janhunen, and Jari Iinatti. 2017. Performance of alow-power wide-area network based on LoRa technology: Doppler robustness, scalability, and coverage. InternationalJournal of Distributed Sensor Networks 13, 3 (2017), 1550147717699412.

[42] Juha Petäjäjärvi, Konstantin Mikhaylov, Antti Roivainen, Tuomo Hanninen, and Marko Pettissalo. 2015. On thecoverage of LPWANs: range evaluation and channel attenuation model for LoRa technology. In ITS Telecommunications(ITST), 2015 14th International Conference on. IEEE, 55–59.


https://github.com/TheThingsNetwork/packet_forwarder

https://github.com/TheThingsNetwork/packet_forwarder

https://doi.org/10.4108/eai.26-10-2015.150597

https://doi.org/10.1109/INFOCOM.2018.8486301

https://doi.org/10.1109/INFOCOM.2018.8486301

https://doi.org/10.1109/COA.2016.7535703

https://doi.org/10.1109/UkrMiCo.2016.7739601

https://doi.org/10.1109/UkrMiCo.2016.7739601


[43] Juha Petäjäjärvi, Konstantin Mikhaylov, Rumana Yasmin, Matti Hämäläinen, and Jari Iinatti. 2017. Evaluation of LoRaLPWAN technology for indoor remote health and wellbeing monitoring. International Journal of Wireless InformationNetworks 24, 2 (2017), 153–165.

[44] Tara Petrić, Mathieu Goessens, Loutfi Nuaymi, Laurent Toutain, and Alexander Pelov. 2016. Measurements, perfor-mance and analysis of LoRa FABIAN, a real-world implementation of LPWAN. In Personal, Indoor, and Mobile RadioCommunications (PIMRC), 2016 IEEE 27th Annual International Symposium on. IEEE, 1–7.

[45] Congduc Pham. 2018. Investigating and experimenting CSMA channel access mechanisms for LoRa IoT networks. InWireless Communications and Networking Conference (WCNC), 2018 IEEE. IEEE, 1–6.

[46] Power monitor 2014. Mobile Device Power Monitor Manual. Rev. 1.14.[47] Brecht Reynders, Wannes Meert, and Sofie Pollin. 2017. Power and spreading factor control in low power wide area

networks. In Communications (ICC), 2017 IEEE International Conference on. IEEE, 1–6.[48] Pieter Robyns, Peter Quax, Wim Lamotte, and William Thenaers. 2018. A Multi-Channel Software Decoder for the

LoRa Modulation Scheme. In Proceedings of the 3rd International Conference on Internet of Things, Big Data and Security- Volume 1: IoTBDS. INSTICC, SciTePress, 41–51.

[49] Tom Rondeau, Josh Blum, Johnathan Corgan, Eric Blossom, and Sebastian Koslowski. [n. d.]. GitHub - gnura-dio/gnuradio: GNU Radio. https://github.com/gnuradio/gnuradio

[50] Olivier Bernard André Seller. 2017. Wireless communication method. US Patent 9,647,718.[51] Olivier Bernard André Seller and Nicolas Sornin. 2016. Low power long range transmitter. US Patent 9,252,834.[52] Semtech. 2012. Semtech Acquires Wireless Long Range IP Provider Cycleo | Semtech Corporation. Retrieved Aug 8,

2018 from http://investors.semtech.com/releasedetail.cfm?ReleaseID=655335[53] Longfei Shangguan, Zheng Yang, Alex X Liu, Zimu Zhou, and Yunhao Liu. 2017. STPP: Spatial-temporal phase

profiling-based method for relative RFID tag localization. IEEE/ACM Transactions on Networking 25, 1 (2017), 596–609.[54] Sigfox. 2018. Sigfox - The Global Communications Service Provider for the Internet of Things (IoT). Retrieved Aug 8,

2018 from https://www.sigfox.com/[55] Rashmi Sharan Sinha, Yiqiao Wei, and Seung-Hoon Hwang. 2017. A survey on LPWA technology: LoRa and NB-IoT.

Ict Express 3, 1 (2017), 14–21.[56] Andreas Springer, Mario Huemer, Leonhard Reindl, Clemens CW Ruppel, Alfred Pohl, Franz Seifert, Wolfgang Gugler,

and Robert Weigel. 1998. A robust ultra-broad-band wireless communication system using SAW chirped delay lines.IEEE transactions on microwave theory and techniques 46, 12 (1998), 2213–2219.

[57] SX1261 2017. SX1261/2 Long Range, Low Power, sub-GHz RF Transceiver. Rev. 1.1.[58] SX1276 2016. SX1276/77/78/79 - 137 MHz to 1020 MHz Low Power Long Range Transceiver. Rev. 5.[59] SX1301 2017. SX1301 Datasheet. Ver. 2.3.[60] Vamsi Talla, Mehrdad Hessar, Bryce Kellogg, Ali Najafi, Joshua R Smith, and Shyamnath Gollakota. 2017. Lora

backscatter: Enabling the vision of ubiquitous connectivity. Proceedings of the ACM on Interactive, Mobile, Wearableand Ubiquitous Technologies 1, 3 (2017), 105.

[61] USRP N210 [n. d.]. USRP™ N200/N210 NETWORKED SERIES.[62] Nuttakit Vatcharatiansakul, Panwit Tuwanut, and Chotipat Pornavalai. 2017. Experimental performance evaluation of

LoRaWAN: A case study in Bangkok. In Computer Science and Software Engineering (JCSSE), 2017 14th InternationalJoint Conference on. IEEE, 1–4.

[63] Benny Vejlgaard, Mads Lauridsen, Huan Nguyen, István Z Kovács, Preben Mogensen, and Mads Sorensen. 2017.Coverage and capacity analysis of sigfox, lora, gprs, and nb-iot. In Proceedings of the 2017 IEEE 85th Vehicular TechnologyConference (VTC Spring), Sydney, Australia. 4–7.

[64] Weightless. 2018. Weightless - Setting the Standard for IoT. Retrieved Aug 8, 2018 from http://www.weightless.org/[65] Peng Zhang and Hao Liu. 2006. An ultra-wide band system with chirp spread spectrum transmission technique. In ITS

Telecommunications Proceedings, 2006 6th International Conference on. IEEE, 294–297.


https://github.com/gnuradio/gnuradio

http://investors.semtech.com/releasedetail.cfm?ReleaseID=655335

https://www.sigfox.com/

http://www.weightless.org/

Documents

Known and Unknown Facts of LoRa: Experiences from a …Known and Unknown Facts of LoRa: Experiences from a Large Scale Measurement Study JANSEN C. LIANDO, AMALINDA GAMAGE, AGUSTINUS